Add Infinigen-only fine-tune checkpoints (epochs 2/5/10) and val88 scores

README.md

@@ -134,3 +134,28 @@ Compared with ScanNet val88: objects are farther from the camera (median 2.68 m
 higher (median 1.62 m vs 1.47 m above the floor), the field of view is wider (median 67° vs 58°), and the category mix
 differs (chair 3 % of tracks vs 30 %). In a fine-tuning test, a model trained only on this data shifted its distance
 prior and placed ScanNet objects 15–25 % farther; mixing it with real data is recommended.
+
+## Checkpoints
+
+`checkpoints/infinigen_10ep_lr1e-5/` holds SpatialEncoder v2 Small (Lite) fine-tuned on this dataset only (train split),
+starting from the e100 checkpoint (`SpatialCoding/Data_verification`, `se_small_lite_wds518_e100/checkpoint.pt`; these
+weights derive from it and follow its terms).
+
+- 10 epochs × 1,600 clips per GPU on 8 GPUs (16,000 steps), batch 1 clip per GPU, 518 px, AdamW, learning rate
+  1e-5 → 1e-7 cosine (spatial-encoder group 1e-7 → 1e-9), same losses and augmentation as e100.
+- `checkpoint_epoch{2,5,10}.pt` are full trainer checkpoints (`model`, `optimizer`, `epoch`, ...), 2.4 GB each;
+  `config.yaml` is the resolved training config; `val88_metrics.json` has the scores below.
+- Trained on the first version of the scene JSONs, where `camera_box_2d_rend` was the visible-mask box (now
+  `camera_box_2d_visible`).
+
+ScanNet val88 (GT 2D-box prompts):
+
+| Checkpoint | [email protected] | [email protected] | [email protected] | 3D IoU |
+|---|---|---|---|---|
+| e100 (start, not in this repo) | 51.80 | 8.82 | 78.10 | 0.258 |
+| epoch 2 | 40.20 | 4.98 | 71.98 | 0.222 |
+| epoch 5 | 42.24 | 6.13 | 71.41 | 0.225 |
+| epoch 10 | 41.99 | 5.64 | 70.92 | 0.223 |
+
+The drop comes mainly from the model's distance prior shifting toward this dataset's farther objects (see above); the
+same checkpoints improve sharply on this dataset's val split (rotation error 32° → 14°, centre error 0.67 → 0.41 m).

checkpoints/infinigen_10ep_lr1e-5/checkpoint_epoch10.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:44a11cf6bbe019d4ff5f0d404478d97a469b3cab8dd05d0b46824b542c5f74c5
+size 2374720676

checkpoints/infinigen_10ep_lr1e-5/checkpoint_epoch2.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aec69d083dba628fa527093b8507287c163c8064e93c7ddee1df35ff1fe14cbc
+size 2374720676

checkpoints/infinigen_10ep_lr1e-5/checkpoint_epoch5.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ecbcda90f04af90365838ddf8d940ca46f24ea14b1cb1529b85d4fa6c7be023e
+size 2374720676

checkpoints/infinigen_10ep_lr1e-5/config.yaml

@@ -0,0 +1,453 @@
+paths:
+  dump_file_name: train_data_scaling_lite
+  experiment_log_dir: /weka/home-wei/project-assets/spatial_encoder/qi_verification/finetune/infinigen_10ep_lr1e-5
+  ca1m_train_meta: /weka/home-wei/project-assets/spatial_encoder/dataset/infinigen_traj/pickle
+  ytvis_dir: /weka/home-wei/project-assets/spatial_encoder/dataset/infinigen_traj/CA-1M
+  ca1m_val_meta: /weka/home-wei/project-assets/spatial_encoder/dataset/infinigen_traj/dataset
+  sam3_checkpoint: /weka/home-wei/project-assets/spatial_encoder/qi_verification/se_small_lite_wds518_e100/checkpoint.pt
+  data_storage_path: /weka/home-wei/project-assets/spatial_encoder/dataset/infinigen_traj
+  hyperism_meta_path: /weka/home-wei/project-assets/spatial_encoder/dataset/infinigen_traj/dataset
+  adt_meta_path: /weka/home-wei/project-assets/spatial_encoder/dataset/infinigen_traj/dataset
+scratch:
+  resolution: 518
+  train_norm_mean:
+  - 0.485
+  - 0.456
+  - 0.406
+  train_norm_std:
+  - 0.229
+  - 0.224
+  - 0.225
+  train_batch_size: 1
+  num_train_workers: 16
+  samples_per_worker_per_epoch: 100
+  max_data_epochs: 10
+  hybrid_repeats: 1
+  video_split_per_frame: 15
+  frame_num_range:
+  - 2
+  - 8
+  frame_sample_gap_range:
+  - 1
+  - 5
+  max_num_objects: 12
+  min_crop_visible_ratio: 0.2
+  norm_scale: 2.5
+  val_video_split_per_frame: 8
+  val_frame_num_range:
+  - 8
+  - 8
+  val_frame_sample_gap_range:
+  - 1
+  - 1
+  val_max_num_objects: 10
+  lr: 1.0e-05
+  vision_lr: 1.0e-07
+  weight_decay: 0.1
+  video_transforms_train:
+  - _target_: sam3.train.transforms.basic_for_api.ComposeAPI
+    transforms:
+    - _target_: sam3.train.transforms.basic_for_api.RandomGrayscale
+      p: 0.05
+      consistent_transform: true
+    - _target_: sam3.train.transforms.basic_for_api.ToTensorAPI
+    - _target_: sam3.train.transforms.basic_for_api.NormalizeAPI
+      mean:
+      - 0.485
+      - 0.456
+      - 0.406
+      std:
+      - 0.229
+      - 0.224
+      - 0.225
+  video_transforms_val:
+  - _target_: sam3.train.transforms.basic_for_api.ComposeAPI
+    transforms:
+    - _target_: sam3.train.transforms.basic_for_api.ToTensorAPI
+    - _target_: sam3.train.transforms.basic_for_api.NormalizeAPI
+      mean:
+      - 0.485
+      - 0.456
+      - 0.406
+      std:
+      - 0.229
+      - 0.224
+      - 0.225
+trainer:
+  _target_: sam3.train.trainer.Trainer
+  mode: train_only
+  skip_saving_ckpts: false
+  empty_gpu_mem_cache_after_eval: true
+  skip_first_val: true
+  max_epochs: 10
+  accelerator: cuda
+  seed_value: 123
+  val_epoch_freq: 9999
+  gradient_accumulation_steps: 1
+  cuda:
+    cudnn_benchmark: false
+  distributed:
+    backend: nccl
+    find_unused_parameters: true
+    gradient_as_bucket_view: true
+  model:
+    _target_: sam3.model_builder.build_sam3_multiplex_joint_video_train_model
+    image_size: 518
+    load_from_HF: false
+    checkpoint_path: /weka/home-wei/project-assets/spatial_encoder/qi_verification/se_small_lite_wds518_e100/checkpoint.pt
+    strict_state_dict_loading: false
+    multiplex_count: 1
+    use_fa3: false
+    use_act_checkpoint_multiplex_transformer: true
+    spatial_encoder:
+      _target_: sam3.model_builder.build_lite_spatial_encoder
+      spatial_name: depthanythingv3_lite
+      is_pretrained: false
+      is_encoder_frozen: false
+      is_head_frozen: false
+      img_size: 518
+      pretrained_path: null
+      base_dim: 32
+      feature_dim: 128
+      temporal_kernel_size: 3
+    enable_spatial_fusion: true
+    predict_depth: true
+    use_decoder_spatial_control: false
+    bbox_token_fusion: none
+    bbox_head_mode: reference_per_candidate
+    freeze_detector: true
+    freeze_tracker_backbone: true
+    freeze_tracker_neck: true
+    freeze_tracker_transformer: true
+    freeze_tracker_maskmem_backbone: true
+    freeze_tracker_sam_prompt_encoder: true
+    spatial_encoder_is_encoder_frozen: false
+    spatial_encoder_is_head_frozen: false
+    enable_detector_fusion: false
+    detector_merge_mode: max
+    detector_blend_alpha: 0.5
+    detector_score_threshold: 0.0
+    prob_to_use_pt_input_for_train: 0.5
+    prob_to_use_pt_input_for_eval: 1.0
+    prob_to_use_box_center_input_for_train: 0.5
+    prob_to_use_box_center_input_for_eval: 1.0
+    num_frames_to_correct_for_train: 2
+    num_frames_to_correct_for_eval: 1
+    rand_frames_to_correct_for_train: true
+    rand_frames_to_correct_for_eval: false
+    add_all_frames_to_correct_as_cond: true
+    num_init_cond_frames_for_train: 2
+    rand_init_cond_frames_for_train: true
+    num_init_cond_frames_for_eval: 1
+    rand_init_cond_frames_for_eval: false
+    enable_dynamic_training: false
+    rand_num_transition_points: false
+    max_num_transition_points: 0
+    add_all_transition_frames_as_cond: false
+    num_correction_pt_per_frame: 3
+    pt_sampling_for_eval: center
+    iou_prediction_use_sigmoid: true
+    sigmoid_scale_for_mem_enc: 20.0
+    sigmoid_bias_for_mem_enc: -10.0
+    use_signed_tpos_enc_to_obj_ptrs: true
+    only_obj_ptrs_in_the_past_for_eval: true
+    max_cond_frames_in_attn: -1
+    save_image_features: true
+    use_maskmem_tpos_v2: false
+    use_linear_no_obj_ptr: false
+    memory_encode_chunk_size: 1
+    detach_memory_encoder_inputs: true
+    build_detector: false
+  loss:
+    all:
+      _target_: sam3.train.loss.sam3_loss.VideoMultiStepDepthLoss
+      weight_dict:
+        loss_2d_bbox_iou: 10.0
+        loss_2d_bbox_l1: 5.0
+        loss_3d_bbox_center_in_2d_l1: 50.0
+        loss_3d_bbox_depth: 10.0
+        loss_3d_bbox_dim: 10.0
+        loss_3d_bbox_alpha: 10.0
+        loss_3d_bbox_quat: 10.0
+        loss_3d_bbox_ious: 1.0
+        loss_3d_chamfer: 10.0
+        loss_depth_map: 1.0
+        loss_depth_gradient: 1.0
+        loss_camera_pose: 1.0
+        loss_class: 10.0
+        loss_3d_bbox_l1: 10.0
+        loss_3d_agg_center_in_first_frame_l1: 10.0
+        loss_3d_agg_dim_l1: 10.0
+        loss_3d_agg_quat: 10.0
+        loss_3d_agg_chamfer: 10.0
+        loss_3d_agg_consistency: 1.0
+      depth_sil_coefficient: 10.0
+      focal_alpha: 0.25
+      focal_gamma: 2.0
+      supervise_all_iou: true
+      iou_use_l1_loss: true
+      pred_obj_scores: true
+      bbox_candidate_selection: pred_iou_times_conf
+      focal_gamma_obj_score: 0.0
+      focal_alpha_obj_score: -1.0
+      log_iou: true
+      log_depth_error: true
+      log_yaw_error: true
+      log_pose: true
+      resolution: 518
+      normalize_by_num_stages: true
+      average_depth_over_objects: true
+    default:
+      _target_: sam3.train.loss.sam3_loss.DummyLoss
+  data:
+    train:
+      _target_: sam3.train.data.torch_dataset.TorchDataset
+      enable_distributed_sampler: false
+      shuffle: false
+      prefetch_factor: 2
+      dataset:
+        _target_: sam3.train.data.aggregate_training_dataset.AggregateDataset
+        meta_dir: /weka/home-wei/project-assets/spatial_encoder/dataset/infinigen_traj/pickle
+        base_dir:
+          CA-1M: /weka/home-wei/project-assets/spatial_encoder/dataset/infinigen_traj/CA-1M
+          hyperism: /weka/home-wei/project-assets/spatial_encoder/dataset/infinigen_traj/hyperism
+          ADT: /weka/home-wei/project-assets/spatial_encoder/dataset/infinigen_traj/aria_digital_twin/ADT
+        transforms:
+        - _target_: sam3.train.transforms.basic_for_api.ComposeAPI
+          transforms:
+          - _target_: sam3.train.transforms.basic_for_api.RandomGrayscale
+            p: 0.05
+            consistent_transform: true
+          - _target_: sam3.train.transforms.basic_for_api.ToTensorAPI
+          - _target_: sam3.train.transforms.basic_for_api.NormalizeAPI
+            mean:
+            - 0.485
+            - 0.456
+            - 0.406
+            std:
+            - 0.229
+            - 0.224
+            - 0.225
+        load_segmentation: true
+        training: true
+        max_train_queries: 100000
+        default_scale: 1000.0
+        predict_metric: true
+        norm_scale: 2.5
+        video_split_per_frame: 15
+        frame_num_range:
+        - 2
+        - 8
+        frame_sample_gap_range:
+        - 1
+        - 5
+        max_num_objects: 12
+        resolution: 518
+        resolution_aug_scale:
+        - 0.8
+        - 1.2
+        min_crop_visible_ratio: 0.2
+        spatial_type: depthanythingv3
+        spatial_resolution: 518
+        shuffle: true
+        num_workers: 16
+        samples_per_worker_per_epoch: 100
+        meta_dir_dict:
+          CA-1M: /weka/home-wei/project-assets/spatial_encoder/dataset/infinigen_traj/pickle
+          hyperism: /weka/home-wei/project-assets/spatial_encoder/dataset/infinigen_traj/dataset
+          ADT: /weka/home-wei/project-assets/spatial_encoder/dataset/infinigen_traj/dataset
+        default_scales:
+          CA-1M: 1000.0
+          hyperism: 1000.0
+          ADT: 1000.0
+        weights:
+        - 0.4
+        - 0.2
+        - 0.4
+        catalog_path: /weka/home-wei/project-assets/spatial_encoder/qi_verification/finetune/infinigen_catalog.json
+      batch_size: 1
+      num_workers: 16
+      pin_memory: true
+      drop_last: true
+      collate_fn:
+        _target_: sam3.train.data.collator.collate_fn_api
+        _partial_: true
+        repeats: 1
+        dict_key: all
+        with_seg_masks: true
+        enforce_single_datapoint_batch: true
+    val:
+      _target_: sam3.train.data.torch_dataset.TorchDataset
+      prefetch_factor: 2
+      dataset:
+        _target_: sam3.train.data.ca1m_dataset.CA1MValDataset
+        meta_dir: /weka/home-wei/project-assets/spatial_encoder/dataset/infinigen_traj/dataset
+        base_dir: /weka/home-wei/project-assets/spatial_encoder/dataset/infinigen_traj/CA-1M
+        transforms:
+        - _target_: sam3.train.transforms.basic_for_api.ComposeAPI
+          transforms:
+          - _target_: sam3.train.transforms.basic_for_api.ToTensorAPI
+          - _target_: sam3.train.transforms.basic_for_api.NormalizeAPI
+            mean:
+            - 0.485
+            - 0.456
+            - 0.406
+            std:
+            - 0.229
+            - 0.224
+            - 0.225
+        load_segmentation: true
+        training: false
+        max_train_queries: 100000
+        max_val_queries: 100000
+        default_scale: 1000.0
+        predict_metric: true
+        use_extrinsic: true
+        norm_scale: 2.5
+        video_split_per_frame: 8
+        frame_num_range:
+        - 8
+        - 8
+        frame_sample_gap_range:
+        - 1
+        - 1
+        max_num_objects: 10
+        resolution: 518
+        min_crop_visible_ratio: 0.2
+        spatial_type: depthanythingv3
+        spatial_resolution: 518
+      shuffle: false
+      batch_size: 1
+      num_workers: 16
+      pin_memory: true
+      drop_last: false
+      collate_fn:
+        _target_: sam3.train.data.collator.collate_fn_api
+        _partial_: true
+        repeats: 1
+        dict_key: all
+        with_seg_masks: true
+        enforce_single_datapoint_batch: true
+  meters:
+    train: null
+    val:
+      all:
+        miou:
+          _target_: sam3.train.metric.MeanIoU
+        translation_error:
+          _target_: sam3.train.metric.TranslationError
+          scale: 2.5
+        depth_error:
+          _target_: sam3.train.metric.DepthError
+          scale: 2.5
+        yaw_error:
+          _target_: sam3.train.metric.YawError
+          unit: rad
+        scale_error:
+          _target_: sam3.train.metric.ScaleError
+  optim:
+    amp:
+      enabled: true
+      amp_dtype: bfloat16
+    optimizer:
+      _target_: torch.optim.AdamW
+      lr: 1.0e-05
+      weight_decay: 0.1
+    options:
+      lr:
+      - scheduler:
+          _target_: fvcore.common.param_scheduler.CosineParamScheduler
+          start_value: 1.0e-05
+          end_value: 1.0000000000000001e-07
+      - scheduler:
+          _target_: fvcore.common.param_scheduler.CosineParamScheduler
+          start_value: 1.0e-07
+          end_value: 9.999999999999999e-10
+        param_names:
+        - tracker.backbone.*
+        - tracker.spatial_*
+      weight_decay:
+      - scheduler:
+          _target_: fvcore.common.param_scheduler.ConstantParamScheduler
+          value: 0.1
+      - scheduler:
+          _target_: fvcore.common.param_scheduler.ConstantParamScheduler
+          value: 0.0
+        param_names:
+        - '*bias*'
+        module_cls_names:
+        - torch.nn.LayerNorm
+    gradient_clip:
+      _target_: sam3.train.optim.optimizer.GradientClipper
+      max_norm: 0.1
+      norm_type: 2
+  checkpoint:
+    save_dir: /weka/home-wei/project-assets/spatial_encoder/qi_verification/finetune/infinigen_10ep_lr1e-5/checkpoints
+    save_freq: 0
+    save_list:
+    - 2
+    - 5
+    - 10
+  logging:
+    tensorboard_writer:
+      _target_: sam3.train.utils.logger.make_tensorboard_logger
+      log_dir: /weka/home-wei/project-assets/spatial_encoder/qi_verification/finetune/infinigen_10ep_lr1e-5/tensorboard
+      flush_secs: 120
+      should_log: true
+    wandb_writer:
+      _target_: sam3.train.utils.wandb_logger.WandbLogger
+      enabled: false
+      project: spatial-encoder
+      entity: null
+      name: train_data_scaling_lite
+      run_id: null
+      mode: disabled
+      log_dir: /weka/home-wei/project-assets/spatial_encoder/qi_verification/finetune/infinigen_10ep_lr1e-5/wandb
+      config:
+        resolution: 518
+        spatial_encoder:
+          _target_: sam3.model_builder.build_lite_spatial_encoder
+          spatial_name: depthanythingv3_lite
+          is_pretrained: false
+          is_encoder_frozen: false
+          is_head_frozen: false
+          img_size: 518
+          pretrained_path: null
+          base_dim: 32
+          feature_dim: 128
+          temporal_kernel_size: 3
+        batch_size_per_gpu: 1
+        num_workers_per_gpu: 16
+        learning_rate: 1.0e-05
+        frame_num_range:
+        - 2
+        - 8
+        gradient_accumulation_steps: 1
+        distributed:
+          backend: nccl
+          find_unused_parameters: true
+          gradient_as_bucket_view: true
+        memory_encode_chunk_size: 1
+        activation_checkpointing: true
+        freeze_inactive_lite_adapters: false
+        workload_window_size: 0
+        workload_window_sort: true
+    log_performance: true
+    log_dir: /weka/home-wei/project-assets/spatial_encoder/qi_verification/finetune/infinigen_10ep_lr1e-5/logs/
+    log_freq: 10
+    log_scalar_frequency: 10
+launcher:
+  num_nodes: 1
+  gpus_per_node: 8
+  experiment_log_dir: /weka/home-wei/project-assets/spatial_encoder/qi_verification/finetune/infinigen_10ep_lr1e-5
+submitit:
+  account: null
+  partition: null
+  qos: null
+  timeout_hour: 72
+  use_cluster: false
+  cpus_per_task: 10
+  port_range:
+  - 45000
+  - 65000
+  constraint: null

checkpoints/infinigen_10ep_lr1e-5/val88_metrics.json

@@ -0,0 +1,49 @@
+{
+ "benchmark": "ScanNet val88 (VSIBench), GT 2D-box prompts, spatial_encoder_v2_small training/final_perception.py protocol",
+ "results": {
+  "e100": {
+   "[email protected]": 0.7810457516339869,
+   "[email protected]": 0.5179738562091504,
+   "[email protected]": 0.08823529411764706,
+   "iou_3d": 0.25839499509443703,
+   "depth_error": 0.14577651820532492,
+   "scale_error": 0.5704675641607397,
+   "translation_error": 0.3705654483154456,
+   "gt_num": 1224,
+   "pred_num": 1224
+  },
+  "epoch2": {
+   "[email protected]": 0.7197712418300652,
+   "[email protected]": 0.40196078431372545,
+   "[email protected]": 0.049836601307189546,
+   "iou_3d": 0.22210775781804434,
+   "depth_error": 0.18222831868170564,
+   "scale_error": 0.6448121021588012,
+   "translation_error": 0.4404954523341529,
+   "gt_num": 1224,
+   "pred_num": 1224
+  },
+  "epoch5": {
+   "[email protected]": 0.7140522875816993,
+   "[email protected]": 0.4223856209150327,
+   "[email protected]": 0.061274509803921566,
+   "iou_3d": 0.22525062597143894,
+   "depth_error": 0.19174793674208948,
+   "scale_error": 0.62655902652468,
+   "translation_error": 0.4638944240397198,
+   "gt_num": 1224,
+   "pred_num": 1224
+  },
+  "epoch10": {
+   "[email protected]": 0.7091503267973857,
+   "[email protected]": 0.4199346405228758,
+   "[email protected]": 0.056372549019607844,
+   "iou_3d": 0.22261301632950997,
+   "depth_error": 0.19339099277919544,
+   "scale_error": 0.6341708121626839,
+   "translation_error": 0.4676733579358411,
+   "gt_num": 1224,
+   "pred_num": 1224
+  }
+ }
+}